As part of today’s International Supercomputing 2021 (ISC) announcements, Intel is showcasing that it will be launching a version of its upcoming Sapphire Rapids (SPR) Xeon Scalable processor with high-bandwidth memory (HBM). This version of SPR-HBM will come later in 2022, after the main launch of Sapphire Rapids, and Intel has stated that it will be part of its general availability offering to all, rather than a vendor-specific implementation.

Hitting a Memory Bandwidth Limit

As core counts have increased in the server processor space, the designers of these processors have to ensure that there is enough data for the cores to enable peak performance. This means developing large fast caches per core so enough data is close by at high speed, there are high bandwidth interconnects inside the processor to shuttle data around, and there is enough main memory bandwidth from data stores located off the processor.


Our Ice Lake Xeon Review system with 32 DDR4-3200 Slots

Here at AnandTech, we have been asking processor vendors about this last point, about main memory, for a while. There is only so much bandwidth that can be achieved by continually adding DDR4 (and soon to be DDR5) memory channels. Current eight-channel DDR4-3200 memory designs, for example, have a theoretical maximum of 204.8 gigabytes per second, which pales in comparison to GPUs which quote 1000 gigabytes per second or more. GPUs are able to achieve higher bandwidths because they use GDDR, soldered onto the board, which allows for tighter tolerances at the expense of a modular design. Very few main processors for servers have ever had main memory be integrated at such a level.


Intel Xeon Phi 'KNL' with 8 MCDRAM Pads in 2015

One of the processors that used to be built with integrated memory was Intel’s Xeon Phi, a product discontinued a couple of years ago. The basis of the Xeon Phi design was lots of vector compute, controlled by up to 72 basic cores, but paired with 8-16 GB of on-board ‘MCDRAM’, connected via 4-8 on-board chiplets in the package. This allowed for 400 gigabytes per second of cache or addressable memory, paired with 384 GB of main memory at 102 gigabytes per second. However, since Xeon Phi was discontinued, no main server processor (at least for x86) announced to the public has had this sort of configuration.

New Sapphire Rapids with High-Bandwidth Memory

Until next year, that is. Intel’s new Sapphire Rapids Xeon Scalable with High-Bandwidth Memory (SPR-HBM) will be coming to market. Rather than hide it away for use with one particular hyperscaler, Intel has stated to AnandTech that they are committed to making HBM-enabled Sapphire Rapids available to all enterprise customers and server vendors as well. These versions will come out after the main Sapphire Rapids launch, and entertain some interesting configurations. We understand that this means SPR-HBM will be available in a socketed configuration.

Intel states that SPR-HBM can be used with standard DDR5, offering an additional tier in memory caching. The HBM can be addressed directly or left as an automatic cache we understand, which would be very similar to how Intel's Xeon Phi processors could access their high bandwidth memory.

Alternatively, SPR-HBM can work without any DDR5 at all. This reduces the physical footprint of the processor, allowing for a denser design in compute-dense servers that do not rely much on memory capacity (these customers were already asking for quad-channel design optimizations anyway).

The amount of memory was not disclosed, nor the bandwidth or the technology. At the very least, we expect the equivalent of up to 8-Hi stacks of HBM2e, up to 16GB each, with 1-4 stacks onboard leading to 64 GB of HBM. At a theoretical top speed of 460 GB/s per stack, this would mean 1840 GB/s of bandwidth, although we can imagine something more akin to 1 TB/s for yield and power which would still give a sizeable uplift. Depending on demand, Intel may fill out different versions of the memory into different processor options.

One of the key elements to consider here is that on-package memory will have an associated power cost within the package. So for every watt that the HBM requires inside the package, that is one less watt for computational performance on the CPU cores. That being said, server processors often do not push the boundaries on peak frequencies, instead opting for a more efficient power/frequency point and scaling the cores. However HBM in this regard is a tradeoff - if HBM were to take 10-20W per stack, four stacks would easily eat into the power budget for the processor (and that power budget has to be managed with additional controllers and power delivery, adding complexity and cost).

One thing that was confusing about Intel’s presentation, and I asked about this but my question was ignored during the virtual briefing, is that Intel keeps putting out different package images of Sapphire Rapids. In the briefing deck for this announcement, there was already two variants. The one above (which actually looks like an elongated Xe-HP package that someone put a logo on) and this one (which is more square and has different notches):

There have been some unconfirmed leaks online showcasing SPR in a third different package, making it all confusing.

 

Sapphire Rapids: What We Know

Intel has been teasing Sapphire Rapids for almost two years as the successor to its Ice Lake Xeon Scalable family of processors. Built on 10nm Enhanced SuperFin, SPR will be Intel’s first processors to use DDR5 memory, have PCIe 5 connectivity, and support CXL 1.1 for next-generation connections. Also on memory, Intel has stated that Sapphire Rapids will support Crow Pass, the next generation of Intel Optane memory.

For core technology, Intel (re)confirmed that Sapphire Rapids will be using Golden Cove cores as part of its design. Golden Cove will be central to Intel's Alder Lake consumer processor later this year, however Intel was quick to point out that Sapphire Rapids will offer a ‘server-optimized’ configuration of the core. Intel has done this in the past with both its Skylake Xeon and Ice Lake Xeon processors wherein the server variant often has a different L2/L3 cache structure than the consumer processors, as well as a different interconnect (ring vs mesh, mesh on servers).

Sapphire Rapids will be the core processor at the heart of the Aurora supercomputer at Argonne National Labs, where two SPR processors will be paired with six Intel Ponte Vecchio accelerators, which will also be new to the market. Today's announcement confirms that Aurora will be using the SPR-HBM version of Sapphire Rapids.

As part of this announcement today, Intel also stated that Ponte Vecchio will be widely available, in OAM and 4x dense form factors:

Sapphire Rapids will also be the first Intel processors to support Advanced Matrix Extensions (AMX), which we understand to help accelerate matrix heavy workflows such as machine learning alongside also having BFloat16 support. This will be paired with updates to Intel’s DL Boost software and OneAPI support. As Intel processors are still very popular for machine learning, especially training, Intel wants to capitalize on any future growth in this market with Sapphire Rapids. SPR will also be updated with Intel’s latest hardware based security.

It is highly anticipated that Sapphire Rapids will also be Intel’s first multi compute-die Xeon where the silicon is designed to be integrated (we’re not counting Cascade Lake-AP Hybrids), and there are unconfirmed leaks to suggest this is the case, however nothing that Intel has yet verified.

The Aurora supercomputer is expected to be delivered by the end of 2021, and is anticipated to not only be the first official deployment of Sapphire Rapids, but also SPR-HBM. We expect a full launch of the platform sometime in the first half of 2022, with general availability soon after. The exact launch of SPR-HBM beyond HPC workloads is unknown, however given those time frames, Q4 2022 seems fairly reasonable depending on how aggressive Intel wants to attack the launch in light of any competition from other x86 vendors or Arm vendors. Even with SPR-HBM being offered to everyone, Intel may decide to prioritize key HPC customers over general availability.

Related Reading

Comments Locked

149 Comments

View All Comments

  • raddude9 - Thursday, July 1, 2021 - link

    HBM, is very power efficient, as you say, because it is basically multiple DRAM dies that are stacked on top of each other, but to avoid overheating, the memory has to run much slower that regular DRAM. HBM compensates for the slow memory speed by having a very wide interface, but there is no way to get around the increased memory latency that having slower RAM will entail

    As for AMD's approach. AMD only announced details of adding a single additional layer of SRAM to it's chips. TSMC on the other hand, has said that this approach can be used to stack up to 8 additional layers of SRAM.
  • mode_13h - Friday, July 2, 2021 - link

    > HBM compensates for the slow memory speed by having a very wide interface

    I thought it was just that they used a wider interface because they *could*. Since it's in-package, there are multiple reasons it's practical to have a much wider interface than external DRAM.

    As for the width vs. frequency, I thought that was just about the interface - not the memory, itself.

    > there is no way to get around the increased memory latency that having slower RAM will entail

    Do you have any latency figures on HBM2? It's be nice to see how it compares with DDR4 and DDR5.

    > TSMC on the other hand, has said that this approach can be used to stack up to
    > 8 additional layers of SRAM.

    I thought there was some discussion that, because the signals were traveling from silicon-to-silicon (i.e. not through TSVs), it was only good for a second layer, but the article definitely suggests more layers could be possible. I still wonder about heat, if going beyond that.
  • raddude9 - Saturday, July 3, 2021 - link

    > I still wonder about heat, if going beyond that.

    stacked SRAM has some major advantages compered to stacked DRAM when it come to heat. Firstly it's a lot less dense than DRAM, and second, it's static, it does not need constant refreshing like DRAM. Sure it takes more power when it comes to fetching memory, but that's not such a big problem.
  • mode_13h - Sunday, July 4, 2021 - link

    > Firstly it's a lot less dense than DRAM

    That's not transistor density, right? That's just cell density, because each cell requires more transistors.

    > it's static, it does not need constant refreshing like DRAM.

    This is an interesting point. How much of DRAM's idle power is due to refreshes? And how does its active power compare with idle power?

    > Sure it takes more power when it comes to fetching memory

    Thing is, L3 cache sees a ton of activity, compared to a DRAM die. AMD was claiming like 40x the bandwidth of DRAM.
  • raddude9 - Monday, July 5, 2021 - link

    > That's not transistor density, right? That's just cell density, because each cell requires more transistors.

    Yes, SRAM cell density is lower

    > Thing is, L3 cache sees a ton of activity, compared to a DRAM die. AMD was claiming like 40x the bandwidth of DRAM.

    Yes indeed, but you were concerned about power usage, if you stack SRAM you end up distributing the power usage over a large area (lower cell density again) and over multiple levels, so heat is probably even less of a concern than it would be for a single layer of SRAM.
  • mode_13h - Wednesday, June 30, 2021 - link

    Another point about scalability is that AMD's stacked SRAM is going to put a huge amount of strain on their interconnect fabric. It will work very well as essentially an extension to a die's L2 cache, but won't help other dies nearly as much.

    I can already tell that AnandTech is going to have fun benchmarking multi-die CPUs with this stacked SRAM, especially if AMD puts it in an EPYC or Threadripper!
  • JayNor - Wednesday, June 30, 2021 - link

    Intel included stacked SRAM on their Ponte Vecchio chip, which Raja called Rambo Cache. They said the HBM was too slow. We should get more detail from Hotchips, but perhaps Anandtech already has the presentations as part of the early release privilege.

    They also announced a hybrid bonding stacked sram that was back in the lab last August. I haven't seen any announcement whether or not the hybrid bonding version is the same as the Rambo Cache.
  • mode_13h - Thursday, July 1, 2021 - link

    > Intel included stacked SRAM on their Ponte Vecchio chip

    I remember the term Rambo Cache, but either missed or forgot that detail. Thanks.
  • webdoctors - Wednesday, June 30, 2021 - link

    Is the HBM supposed to replace the DRAM in the hierarchy?

    "leading to 64 GB of HBM."

    That seems far too little for servers that would have 100s of GBs of main memory.
  • mode_13h - Thursday, July 1, 2021 - link

    Or even TBs. Yeah, it can supposedly be used as L4 cache.

Log in

Don't have an account? Sign up now